CSSG: A Continuous Spatio-temporal Graph Learning Framework with Scalable Spatial Granularity

Public Health Emergencies (PHEs) pose serious threats to public health and cause socio-economic disruptions. During such events, mobility restrictions lead to a surge in online shopping for daily supplies. Predicting multi-category product demand under PHEs is essential to help e-commerce and logistics companies improve the efficiency of material management and regional distribution. However, shifting policy interventions alter both population distribution and consumer intent, making demand prediction particularly challenging. In this paper, we propose M2DP, a novel multi-scale association learning framework for multi-category demand prediction under PHEs. Our approach includes: (1) multi-scale temporal encoders to extract long/short-term demand patterns and an event evolution module to incorporate regional impacts; (2) a dual association learning module that integrates both empirical and adaptive category correlations; and (3) a regional joint prediction module that combines temporal and associative features via a spatial graph for final forecasting. Experiments on real-world e-commerce data show that our model achieves state-of-the-art performance.

Recommended citation: Xia, K., Lin, L., Zhang, Q., Zhang, X., & Wang, S., Hu X. & Yu P (2026, August). CSSG: A Continuous Spatio-temporal Graph Learning Framework with Scalable Spatial Granularity. In Proceedings of the 32st ACM SIGKDD Conference on Knowledge Discovery and Data Mining. 2026.